This study presents a new \(SRL\) model with physically defensible scaling for unbounded and width-limited ruptures.
Study Objectives:
The empirical data that are used in model development include the FDHI (Sarmiento et al. 2021), Wells and Coppersmith (1994), and Baize et al. (2020) datasets. The surface rupture lengths (\(SRL\)) of the events in FDHI dataset are computed based on the Event Coordinate System (ECS), while the \(SRL\) in Wells and Coppersmith (1994) and Baize et al. (2020) are reported directly. Additionally, dynamic rupture simulations performed by SCEC are used to derive the rupture width model for the width-limited ruptures.
#read srl flatfiles
#FDHI
df_FDHI_srl_all <- read.csv(fname_FDHI_srl)
#FDHI processing
#create normal and reverse fault flags
df_FDHI_srl_all$FN <- pmax(-df_FDHI_srl_all$sof, 0)
df_FDHI_srl_all$FR <- pmax( df_FDHI_srl_all$sof, 0)
df_FDHI_srl_all$FD <- abs( df_FDHI_srl_all$sof )
#style of faulting name
df_FDHI_srl_all$style2 <- 'Strike-Slip'
df_FDHI_srl_all[df_FDHI_srl_all$FN>0,'style2'] <- 'Normal'
df_FDHI_srl_all[df_FDHI_srl_all$FR>0,'style2'] <- 'Reverse'
#fault width
df_FDHI_srl_all$seis_width <- df_FDHI_srl_all$seis_depth / sin(deg2rad(df_FDHI_srl_all$dip))
#SCEC
#read srl flatfiles
df_SCEC_srl_original <- read.csv(fname_SCEC_srl_original)
df_SCEC_srl_low_width <- read.csv(fname_SCEC_srl_low_width)
df_SCEC_srl_const_sig <- read.csv(fname_SCEC_srl_const_sig)
df_SCEC_srl_sim_2 <- read.csv(fname_SCEC_srl_sim2)
df_SCEC_srl_sim_3 <- read.csv(fname_SCEC_srl_sim3)
df_SCEC_srl_sim_4 <- read.csv(fname_SCEC_srl_sim4)
#SCEC processing
#define seimogenic depths
df_SCEC_srl_original$seis_depth <- 20
df_SCEC_srl_low_width$seis_depth <- 15
df_SCEC_srl_const_sig$seis_depth <- 20
df_SCEC_srl_sim_2$seis_depth <- 20
df_SCEC_srl_sim_3$seis_depth <- 20
df_SCEC_srl_sim_4$seis_depth <- 20
#merged data-frame of different simulations
df_SCEC_srl <- bind_rows( df_SCEC_srl_original, df_SCEC_srl_low_width, df_SCEC_srl_const_sig,
df_SCEC_srl_sim_2, df_SCEC_srl_sim_3, df_SCEC_srl_sim_4)
#identify width limited events
df_SCEC_srl$flag_wlim <- FALSE
df_SCEC_srl[df_SCEC_srl$wef >= df_SCEC_srl$seis_depth-2.5,'flag_wlim'] <- TRUE
rm(list=c('df_SCEC_srl_original','df_SCEC_srl_low_width','df_SCEC_srl_const_sig',
'df_SCEC_srl_sim_2', 'df_SCEC_srl_sim_3', 'df_SCEC_srl_sim_4'))
The maximum fault width of the empirical data was calculated based on the thickness of the seismogenic zone (\(D_{seis}\)) and fault dip angle (\(\theta\)). Table 1 summarizes \(D_{seis}\) based on the fault location and Table 2 summarizes \(\theta\) for different styles of faulting. \[ W_{lim} = \frac{D_{seis}}{\sin(\theta)} \] \(W_{lim}\) is the maximum width, a rupture can reach before it becomes width limited.
| Region | Seismogenic Depth (km) |
|---|---|
| California | 15 |
| Guatemala | 15 |
| New Zealand | 15 |
| Indonesia | 15 |
| Japan | 17 |
| Himalayan Region | 30 |
| Australia | 40 |
| Other | 22 |
| Style of Fauling | Dip Angle (deg) |
|---|---|
| Strike-Slip | 90 |
| Normal | 60 |
| Reverse | 45 |
The following events were excluded from the regression dataset:
#events for regression
df_FDHI_srl <- df_FDHI_srl_all
#remove outliers
df_FDHI_srl <- subset(df_FDHI_srl, !is.na(event_len) & flag_missing==0 & flag_multiple==0 )
#1892 Laguna Salada
df_FDHI_srl <- subset(df_FDHI_srl, !(EQ_ID==40 & FLT_ID==1))
#1978 IzuOshima
df_FDHI_srl <- subset(df_FDHI_srl, !(EQ_ID==62 & FLT_ID==1))
#1993 Killari
df_FDHI_srl <- subset(df_FDHI_srl, !(EQ_ID==74 & FLT_ID==1))
#2010 Pisayambo
df_FDHI_srl <- subset(df_FDHI_srl, !(EQ_ID==58 & FLT_ID==1))
# #1957 Gobi-Altai
# df_FDHI_srl <- subset(df_FDHI_srl, !(EQ_ID==52 & FLT_ID==2))
#2018 Etna
df_FDHI_srl <- subset(df_FDHI_srl, !(EQ_ID==6 & FLT_ID==3))
#remove mag less than 5
df_FDHI_srl <- subset(df_FDHI_srl, magnitude>=5)
#summarize all outliers
df_FDHI_srl_out <- subset(df_FDHI_srl_all,!(rownames(df_FDHI_srl_all)%in%rownames(df_FDHI_srl)))
Summary statistics of the empirical dataset are presented below:
## Number of Events per dataset:
## Total: 122
## FDHA: 64
## W&C: 52
## Baize: 6
## Number of Events per region:
## Total: 122
## California: 29
## Guatemala: 1
## New Zealand: 3
## Indonesia: 1
## Japan: 7
## Himalayan Region: 12
## Australia: 11
## Other: 58
## Number of Events per Style of Faulting:
## Strike-slip: 64
## Normal: 26
## Reverse: 32
## Magnitude Range:
## maximum: 8.1
## minimum: 5.0
## Rupture-Length Range:
## maximum: 432.0
## minimum: 1.1
Figure 1 shows the Magnitude - Surface rupture length (\(\mathbf{M} - SRL\)) distribution of the empirical dataset.
Figure 1. Magnitude - SRL distribution of selected empirical data
Summary statistics of the SCEC simulations are presented below:
## Total number of simulations: 554
## Magnitude Range:
## maximum: 8.2
## minimum: 4.9
## Rupture-Length Range:
## maximum: 655.2
## minimum: 1.0
Figure 2 compares the \(\mathbf{M} - SRL\) distribution of empirical data and SCEC simulations. The two datasets show similar scaling supporting that the scaling relationships derived from the SCEC simulations are applicable to the empirical data.
Figure 2. Comparison of Magnitude - SRL distribution of empirical data and SCEC simulations
The SCEC simulations were used to develop a Magnitude - Rupture width
model for unbounded and width-limited ruptures. Two models were
evaluated: (i) a linear model between \(W -
\mathbf{M}\), hereafter denoted as model w0, and
(ii) a model with a scaling break at \(W_{lim}\), hereafter denoted as
model w1
The functional form for model_w0 is: \[\begin{equation}
\log(W) = \mathbf(M) + \epsilon_W
\end{equation}\]
The functional form for model_w1 is: \[\begin{equation}
log(W) = \min(\beta_1 + \beta_2 \mathbf{M}, \log(W_{lim})) + \epsilon_W
\end{equation}\]
## Regression formulas
form_model_w0 <- log10(wef) ~ 1 + magnitude
form_model_w1 <- log10(wef) ~ 1 + magnitude
## Regression
fit_model_w0 <- lm(form_model_w0, data=df_SCEC_srl)
fit_model_w1 <- lm(form_model_w1, data=subset(df_SCEC_srl, !flag_wlim))
The \(W - \mathbf{M}\) model summary
is presented below. model w1 has a higher log likelihood
and lower Akaike information criterion.
## Width Modeling
##
## Model0:
## LogLik: 641.71
## AIC: -1277.43
##
## Model1:
## LogLik: 889.34
## AIC: -1772.69
The best estimates of the coefficients of the two candidate models are summarized in Table 3.
| Intercept | \(\mathbf{M}\) scaling | \(\sigma\) | |
|---|---|---|---|
| model_w0 | -0.7603617 | 0.2735708 | 0.0761189 |
| model_w1 | -1.1602315 | 0.3395897 | 0.0544508 |
Figure 3 compares the two candidate models against the SCEC simulations.
## Warning: Using `size` aesthetic for lines was deprecated in ggplot2 3.4.0.
## ℹ Please use `linewidth` instead.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.
Figure 3. Comparison of candidate rupture width models against the SCEC simulations
The summary of the model w1 regression is provided
below:
summary(fit_model_w1)
##
## Call:
## lm(formula = form_model_w1, data = subset(df_SCEC_srl, !flag_wlim))
##
## Residuals:
## Min 1Q Median 3Q Max
## -0.123926 -0.033305 0.000471 0.036243 0.172420
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -1.160232 0.032544 -35.65 <2e-16 ***
## magnitude 0.339590 0.005303 64.04 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.05445 on 328 degrees of freedom
## Multiple R-squared: 0.9259, Adjusted R-squared: 0.9257
## F-statistic: 4101 on 1 and 328 DF, p-value: < 2.2e-16
Figure 4 shows the scaling of the preferred model
(model w1) against the SCEC data.
## Warning: The `size` argument of `element_rect()` is deprecated as of ggplot2 3.4.0.
## ℹ Please use the `linewidth` argument instead.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.
Figure 4. Preferred model for rupture width scaling
Figure 5 shows the residuals of the preferred model versus magnitude.
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
Figure 5. Resisuals of Preferred Rupture Width Model for SCEC simulations
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
This section presents the deviation of the \(SRL\sim\mathbf{M}\) functional form. The Seismic Moment (\(M_0\)) is defined as:
\[\begin{equation} M_0 = \mu~\bar{D}~L~W \end{equation}\]where \(\mu\) is the stiffness of the crust, \(\bar{D}\) is the average slip on the rupture, \(L\) is the length of the rupture, and \(W\) is width of the rupture. Taking the log of the previous equation and substituting the definition of the Moment Magnitude (\(\mathbf{M} = \frac{2}{3}\log(M_0) - 10.7\)) we get:
\[\begin{equation} \mathbf{M} = \frac{2}{3}\log(\mu) + \frac{2}{3}\log(\bar{D}) + \frac{2}{3}\log(L) + \frac{2}{3}\log(W) - 10.7 \end{equation}\]Since stiffness of the crust (\(\mu\)) is constant, the previous equation simplifies to:
\[\begin{equation} \mathbf{M} = \frac{2}{3}\log(\bar{D}) + \frac{2}{3}\log(L) + \frac{2}{3}\log(W) + c \end{equation}\]where \(c\) is used to denote any arbitrary constant.
Three commonly used scaling models for the average displacements
found in the literature include the \(\sqrt{A}\) model, hereafter called
model 1, the \(W\) model,
hereafter called model 2, and the \(L\) model, hereafter called
model 3 (Scholz, 1982; Romanowicz, 1992; Romanowicz and
Rundle, 1993; Romanowicz, 1994; Scholz, 1994; Bodin and Brune, 1996;
Pegler and Das, 1996; Scholz, 1997, 1998; Hanks and Bakun, 2002;
Leonard, 2010).
For model 1, the average displacement is assumed to
scale linearly with the square root of the rupture area:
For model 2, the average displacement is assumed to
scale linearly with the width of the rupture:
while for model 3, the average displacement is assumed
to scale linearly with the length of the rupture:
Using model 1 the \(\mathbf{M}\) scaling relationship becomes:
\[\begin{equation}
\mathbf{M} = \log(L) + \log(W) + c
\end{equation}\] with model 2 the \(\mathbf{M}\) scaling relationship becomes:
\[\begin{equation}
\mathbf{M} = \frac{2}{3}\log(L) + \frac{4}{3}\log(W) + c
\end{equation}\] and with model 3 the \(\mathbf{M}\) scaling relationship becomes:
\[\begin{equation}
\mathbf{M} = \frac{4}{3}\log(L) + \frac{2}{3}\log(W) + c
\end{equation}\]
Solving for \(\log(L)\) in previous
equations model 1 becomes: \[\begin{equation}
\log(L) = \mathbf{M} - \log(W) + c
\end{equation}\] model 2 becomes: \[\begin{equation}
\log(L) = \frac{3}{2} \mathbf{M} - 2 \log(W) + c
\end{equation}\] and model 3 becomes: \[\begin{equation}
\log(L) = \frac{3}{4} \mathbf{M} - \frac{1}{2} \log(W) + c
\end{equation}\]
An additional two candidate models are developed which follow \(\sqrt{A}\) scaling for unbounded ruptures
and transition to \(L\) and \(W\) scaling, respectively, for
width-limited events. The two models are referred to as
model 2' and model 3' and their functional
forms are provided next. The model 2' functional form is:
\[\begin{equation}
\log(L) = \begin{cases} \mathbf{M} - \log(W) + c & \mathbf{M} \le
\mathbf{M}_{lim} \\ \frac{3}{2} \mathbf{M} - 2 \log(W) + c_2' +
c & \mathbf{M} > \mathbf{M}_{lim} \end{cases}
\end{equation}\] while, the model 3' functional form
is: \[\begin{equation}
\log(L) = \begin{cases} \mathbf{M} - \log(W) + c & \mathbf{M} \le
\mathbf{M}_{lim} \\ \frac{3}{4} \mathbf{M} - \frac{1}{2} \log(W) +
c_3' + c & \mathbf{M} > \mathbf{M}_{lim} \end{cases}
\end{equation}\] \(c_2'\)
and \(c_3'\) are added in order to
have a continuous transition between the two branches. The transition
from unbounded to width-limited events occurs at \(\mathbf{M}_{lim}\), \(W_{lim}\) coordinate pair. Solving for
\(c_2'\) and \(c_3'\) at the transition point we
get:
\[\begin{equation}
c_2' = -\frac{1}{2} \mathbf{M}_{lim} + \log(W_{lim})
\end{equation}\] \[\begin{equation}
c_3' = \frac{1}{4} \mathbf{M}_{lim} - \frac{1}{2} \log(W_{lim})
\end{equation}\] With the estimated coefficients, the previous
equations can be simplified. model 2' becomes: \[\begin{equation}
\log(L) = \frac{3}{2} \mathbf{M} - \frac{1}{2}
\min(\mathbf{M},\mathbf{M}_{lim}) - \log(W) + c
\end{equation}\] and model 3' becomes: \[\begin{equation}
\log(L) = \frac{3}{4} \mathbf{M} + \frac{1}{4}
\min(\mathbf{M},\mathbf{M}_{lim}) - \log(W) + c
\end{equation}\]
Substituting the model w1 from the SCEC rupture
simulations to the model 1 functional form we get: \[\begin{equation}
\log(L) = \mathbf{M} - \min(\beta_1 + \beta_2 \mathbf{M}, \log(W_{lim}))
+ c
\end{equation}\] for model 2 we get: \[\begin{equation}
\log(L) = \frac{3}{2} \mathbf{M} - 2 min(\beta_1 + \beta_2 \mathbf{M},
\log(W_{lim})) + c
\end{equation}\] and for model 3 we get: \[\begin{equation}
\log(L) = \frac{3}{4} \mathbf{M} - \frac{1}{2} min(\beta_1 + \beta_2
\mathbf{M}, \log(W_{lim})) + c
\end{equation}\]
Similarly, for model 2' we get: \[\begin{equation}
\log(L) = \frac{3}{2} \mathbf{M} - \frac{1}{2}
\min(\mathbf{M},\mathbf{M}_{lim}) - min(\beta_1 + \beta_2 \mathbf{M},
\log(W_{lim})) + c
\end{equation}\] and for model 3' we get: \[\begin{equation}
\log(L) = \frac{3}{4} \mathbf{M} + \frac{1}{4}
\min(\mathbf{M},\mathbf{M}_{lim}) - min(\beta_1 + \beta_2 \mathbf{M},
\log(W_{lim})) + c
\end{equation}\]
Assuming the surface to sub-surface rupture length scaling is of the
form \(\log(SRL) = c_1 \log(L) + c\),
substituting it to the previous equations, condensing the constants, and
solving for \(\log(SRL)\), for
model 1 we get: \[\begin{equation}
\log(SRL) = c_1 \left( \mathbf{M} - \min(\beta_1 + \beta_2 \mathbf{M},
\log(W_{lim})) \right) + c_2 = c_1 X_1 + c_2
\end{equation}\] for model 2 we get: \[\begin{equation}
\log(SRL) = c_1 \left( \frac{3}{2} \mathbf{M} - 2 \min(\beta_1 + \beta_2
\mathbf{M}, \log(W_{lim})) \right) + c_2 = c_1 X_2 + c_2
\end{equation}\] for model 3 we get: \[\begin{equation}
\log(SRL) = c_1 \left( \frac{3}{4} \mathbf{M} - \frac{1}{2} \min(\beta_1
+ \beta_2 \mathbf{M}, \log(W_{lim})) \right) + c_2 = c_1 X_3 + c_2
\end{equation}\] for model 2' we get: \[\begin{equation}
\log(SRL) = c_1 \left( \frac{3}{2} \mathbf{M} - \frac{1}{2}
\min(\mathbf{M},\mathbf{M}_{lim}) - \min(\beta_1 + \beta_2 \mathbf{M},
\log(W_{lim})) \right) + c_2 = c_1 X_2' + c_2
\end{equation}\] and for model 3' we get: \[\begin{equation}
\log(SRL) = c_1 \left( \frac{3}{4} \mathbf{M} + \frac{1}{4}
\min(\mathbf{M},\mathbf{M}_{lim}) - \min(\beta_1 + \beta_2 \mathbf{M},
\log(W_{lim})) \right) + c_2 = c_1 X_3' + c_2
\end{equation}\]
The previous equations can be expressed as linear models with, \(X_1 = \mathbf{M} - \min(\beta_1 + \beta_2 \mathbf{M}, \log(W_{lim})\), \(X_2 = \frac{3}{2} \mathbf{M} - 2 \min(\beta_1 + \beta_2 \mathbf{M}, \log(W_{lim})\), \(X_3 = \frac{3}{4} \mathbf{M} - \frac{1}{2} \min(\beta_1 + \beta_2 \mathbf{M}, \log(W_{lim})\), \(X_2' = \frac{3}{2} \mathbf{M} - \frac{1}{2} \min(\mathbf{M},\mathbf{M}_{lim}) - min(\beta_1 + \beta_2 \mathbf{M}, \log(W_{lim}))\), and \(X_3' = \frac{3}{4} \mathbf{M} + \frac{1}{4} \min(\mathbf{M},\mathbf{M}_{lim}) - min(\beta_1 + \beta_2 \mathbf{M}, \log(W_{lim}))\).
The effect of the style of faulting is also considered through the
use of additive terms to account for differences in the static stress
drop. The static stress drop is defined as: \[\begin{equation}
\Delta \sigma = C \mu \frac{\bar{D}}{L_C}
\end{equation}\] thus, in log scale, a shift in the stress drop
will be accompanied by a linear shift in the average displacement (\(\log(\bar{D}) = \log(\Delta \sigma) +
const\)), which, in turn, will shift the intercept of the \(SLR \sim \mathbf{M}\) models
Two modeling approaches were evaluated: (i) a common additive term for
dip-slip events (\(F_D\)), and (ii)
separate additive terms for Normal and Reverse faults (\(F_N\) and \(F_R\)). The style of faulting dependence
introduced through \(F_D\), \(F_N\), and \(F_R\) acts in addition to the style of
faulting dependence introduced by the different dip angles between the
different styles of faulting.
The final regression relationship for model 1 is: \[\begin{equation}
\log(SRL) = c_1 X_1 + c_2 F_D + c_3
\end{equation}\] the final regression relationship for
model 2 is: \[\begin{equation}
\log(SRL) = c_1 X_2 + c_2 F_D + c_3
\end{equation}\] the final regression relationship for
model 3 is: \[\begin{equation}
\log(SRL) = c_1 X_3 + c_2 F_D + c_3
\end{equation}\] the final regression relationship for
model 2' is: \[\begin{equation}
\log(SRL) = c_1 X'_2 + c_2 F_D + c_3
\end{equation}\] the final regression relationship for
model 3' is: \[\begin{equation}
\log(SRL) = c_1 X'_3 + c_2 F_D + c_3
\end{equation}\] and the final regression for separate additive
terms with \(D \sim \sqrt{A}\) scaling,
hereafter denoted as model 1alt is: \[\begin{equation}
\log(SRL) = c_1 X_1 + c_2 F_N + c_3 F_R + c_4
\end{equation}\]
Lastly a Wells and Coppersmith type model was also evaluated,
hereafter referred as model 0, to investigate the impact of
width-limited ruptures on current state-of-practice models. \[\begin{equation}
\log(SRL) = c_1 \mathbf{M} + c_2 F_D + c_3
\end{equation}\]
The coefficients of the previous equations are estimated through a maximum likelihood regression.
The seven candidate models evaluated in predicting \(SRL\) as a function of \(\mathbf{M}\) are summarized in Table 4.
| Name | Functional Form | Description |
|---|---|---|
| model 0 | \(\log(SLR)\sim 1 + \mathbf{M} + F_D\) | W&C 94 type model |
| model 1 | \(\log(SLR)\sim1 + X_1 + F_D\) | \(\bar{D} \sim \sqrt{A}\) type model with \(F_D\) addative term |
| model 2 | \(\log(SLR)\sim 1+X_2 + F_D\) | \(\bar{D} \sim {W}\) type model with \(F_D\) addative term |
| model 3 | \(\log(SLR)\sim 1+X_3 + F_D\) | \(\bar{D} \sim {L}\) type model with \(F_D\) addative term |
| model 2’ | \(\log(SLR)\sim 1+X_{2'} + F_D\) | \(\bar{D} \sim \sqrt{A},{W}\) type model with \(F_D\) addative term |
| model 3’ | \(\log(SLR)\sim 1+X_{3'} + F_D\) | \(\bar{D} \sim \sqrt{A},{L}\) type model with \(F_D\) addative term |
| model 1 alt | \(\log(SLR)\sim1 + X_1 + F_N+F_R\) | \(\bar{D} \sim \sqrt{A}\) type model with \(F_N\) & \(F_R\) addative terms |
The next part shows the definition of the regression functional forms and model fit.
# Rupture Width Modeling
#rupture width coefficients
c1_rw <- df_SCEC_model_width_mag_coeffs['model_w1','intrcp']
c2_rw <- df_SCEC_model_width_mag_coeffs['model_w1','M']
#compute rupture widths
df_FDHI_srl$rup_width_unbnd <- 10^(c1_rw + c2_rw*df_FDHI_srl$magnitude)
df_FDHI_srl$rup_width <- pmin(df_FDHI_srl$rup_width_unbnd, df_FDHI_srl$seis_width)
#transitioning magnitude
df_FDHI_srl$magnitude_lim <- 1/c2_rw * (log10(df_FDHI_srl$seis_width) - c1_rw )
#identify width limited events
df_FDHI_srl$flag_wlim <- FALSE
df_FDHI_srl[df_FDHI_srl$rup_width >= df_FDHI_srl$seis_width,'flag_wlim'] <- TRUE
# SRL Modeling
#scaling parameters
#simple models
df_FDHI_srl$x1 <- df_FDHI_srl$magnitude - log10(df_FDHI_srl$rup_width)
df_FDHI_srl$x2 <- 3/2*df_FDHI_srl$magnitude - 2*log10(df_FDHI_srl$rup_width)
df_FDHI_srl$x3 <- 3/4*df_FDHI_srl$magnitude - 1/2*log10(df_FDHI_srl$rup_width)
#composite models
df_FDHI_srl$x2prime <- 3/2*df_FDHI_srl$magnitude - 1/2*pmin(df_FDHI_srl$magnitude,df_FDHI_srl$magnitude_lim) - log10(df_FDHI_srl$rup_width)
df_FDHI_srl$x3prime <- 3/4*df_FDHI_srl$magnitude + 1/4*pmin(df_FDHI_srl$magnitude,df_FDHI_srl$magnitude_lim) - log10(df_FDHI_srl$rup_width)
#standard deviation weights
width_scl <- 0.1
sig_scl <- 2.
df_FDHI_srl$reg_wt <- ( 1 + (sig_scl-1) * sigmoid( 2.197225 * (log10(df_FDHI_srl$rup_width_unbnd)-log10(df_FDHI_srl$seis_width))/ width_scl ) )^2
## Regression formulas
form_model_srl0m <- log10(event_len) ~ 1 + magnitude + FD
form_model_srl1m <- log10(event_len) ~ 1 + x1 + FD
form_model_srl2m <- log10(event_len) ~ 1 + x2 + FD
form_model_srl3m <- log10(event_len) ~ 1 + x3 + FD
form_model_srl2pm <- log10(event_len) ~ 1 + x2prime + FD
form_model_srl3pm <- log10(event_len) ~ 1 + x3prime + FD
form_model_srl1m_alt <- log10(event_len) ~ 1 + x1 + FN + FR
## Regression
fit_model_srl0m <- lm(form_model_srl0m, data=df_FDHI_srl)
fit_model_srl1m <- lm(form_model_srl1m, data=df_FDHI_srl, weights=reg_wt)
fit_model_srl2m <- lm(form_model_srl2m, data=df_FDHI_srl, weights=reg_wt)
fit_model_srl3m <- lm(form_model_srl3m, data=df_FDHI_srl, weights=reg_wt)
fit_model_srl2pm <- lm(form_model_srl2pm, data=df_FDHI_srl, weights=reg_wt)
fit_model_srl3pm <- lm(form_model_srl3pm, data=df_FDHI_srl, weights=reg_wt)
fit_model_srl1m_alt <- lm(form_model_srl1m_alt, data=df_FDHI_srl, weights=reg_wt)
The print out below summarizes the log-likelihood and Akaike Information Criterion (AIC) for the different candidate models, while Table 5 summarizes the estimated coefficients.
From a statistical perspective, the simple models
(model 1 to model 3) and composite
model 2' have similar good predictive performance, with
model 1 being slightly better. The composite
model 3' has poorer performance, while model 0
has the worst performance. model 0 does not consider the
finite thickness of the seismogenic zone showcasing the limitations of a
purely empirical model. The candidate models (model 1 to
model 3') have a break in the magnitude scaling when the
width of the rupture reaches the width of the seismogenic zone allowing
them to better fit the data.
## SRL Modeling:
## Model0:
## LogLik: 6.06
## AIC: -4.12
##
## Model1:
## LogLik: 15.42
## AIC: -22.85
##
## Model2:
## LogLik: 14.15
## AIC: -20.30
##
## Model3:
## LogLik: 14.30
## AIC: -20.60
##
## Model2':
## LogLik: 8.91
## AIC: -9.81
##
## Model3':
## LogLik: 12.22
## AIC: -16.44
##
## Model1 alt.:
## LogLik: 15.43
## AIC: -20.86
## Standard Deviations:
## Model 1:
## Empirical ratio for std of width limted to unbounded events is: 1.93
##
## Standard Deviations:
## Model 2:
## Empirical ratio for std of width limted to unbounded events is: 1.88
##
## Standard Deviations:
## Model 3:
## Empirical ratio for std of width limted to unbounded events is: 1.88
##
## Standard Deviations:
## Model 2':
## Empirical ratio for std of width limted to unbounded events is: 1.74
##
## Standard Deviations:
## Model 3':
## Empirical ratio for std of width limted to unbounded events is: 1.80
##
## Standard Deviations:
## Model 1 alt.:
## Empirical ratio for std of width limted to unbounded events is: 1.92
| Intercept | \(\mathbf{M}\) scaling | \(X_1\) scaling | \(X_2\) scaling | \(X_3\) scaling | \(X_{2'}\) scaling | \(X_{3'}\) scaling | \(F_D\) | \(F_N\) | \(F_R\) | \(\sigma\) | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| model_srl0 | -2.968009 | 0.6628783 | NA | NA | NA | NA | NA | -0.1836268 | NA | NA | 0.2331305 |
| model_srl1 | -3.878287 | NA | 0.9509412 | NA | NA | NA | NA | -0.1517421 | NA | NA | 0.2494025 |
| model_srl2 | -4.179800 | NA | NA | 0.7152433 | NA | NA | NA | -0.1390408 | NA | NA | 0.2520182 |
| model_srl3 | -3.604864 | NA | NA | NA | 1.131573 | NA | NA | -0.1645242 | NA | NA | 0.2517162 |
| model_srl2p | -3.097664 | NA | NA | NA | NA | 0.8044237 | NA | -0.1291246 | NA | NA | 0.2630927 |
| model_srl3p | -4.250663 | NA | NA | NA | NA | NA | 1.023459 | -0.1758600 | NA | NA | 0.2560359 |
| model_srl1_alt | -3.881661 | NA | 0.9515046 | NA | NA | NA | NA | NA | -0.155196 | -0.1480471 | 0.2504442 |
From seismological theory, model 1 to
model 3' should have unit slopes with respect to the
magnitude scaling terms (e.g. \(X_1\),
\(X_2\), \(X_3\)), if the average displacement in the
empirical data follows the assumed scaling law and the subsurface
rupture length (\(L\)) was the response
variable. With \(SRL\) as a response
variable, a magnitude scaling term greater than one would imply that the
surface to sub-surface rupture length ratio increases with magnitude
while a magnitude scaling term less than one would imply that the
surface to sub-surface rupture length ratio decreases with magnitude.
Intuitively, a greater than one magnitude scaling term is expected, as
for large events, a larger part of the rupture is expected to reach the
surface.
Based on the derived slopes, model 1 and
model 3' show a data-derived scaling consistent with
seismological theory, with \(X_1\)
scaling in model 1 being slightly less than unity, and
\(X_3'\) scaling in
model 3' being slightly above unity. Considering the
statistical performance and seismological scaling of the candidate
models, model 1 with a fixed magnitude scaling slope equal
to unity is selected as a preferred model, hereafter called
model1_fxd, and model 3' with fixed unit slope
is denoted as an alternative valid model, hereafter called
model3_fxd'. In the selection process, seismologically
consistent scaling is weighted more heavily than statistical
performance, due to the small size of the empirical data set and
importance of physically plausible model extrapolation seismic hazard
analyses.
## Preferred model
# regression formula
form_model_srl1m_fxd <- log10(event_len) ~ 1 + offset(x1) + FD
form_model_srl3pm_fxd <- log10(event_len) ~ 1 + offset(x3prime) + FD
# regression
fit_model_srl1m_fxd <- lm(form_model_srl1m_fxd, data=df_FDHI_srl, weights=reg_wt)
fit_model_srl3pm_fxd <- lm(form_model_srl3pm_fxd, data=df_FDHI_srl, weights=reg_wt)
#residuals
df_FDHI_srl$srl1m_fxd_res <- resid(fit_model_srl1m_fxd)
df_FDHI_srl$srl3pm_fxd_res <- resid(fit_model_srl3pm_fxd)
#coefficients
df_model_pref_srl_mag_coef <- as.data.frame(t(coef(fit_model_srl1m_fxd)), row.names='model_srl1_fxd (preferred)')
df_model_pref_srl_mag_coef <- bind_rows(df_model_pref_srl_mag_coef, as.data.frame(t(coef(fit_model_srl3pm_fxd)), row.names='model_srl3p_fxd (alternative)'))
df_model_pref_srl_mag_coef['model_srl1_fxd (preferred)', 'x1'] <- 1
df_model_pref_srl_mag_coef['model_srl3p_fxd (alternative)','x3prime'] <- 1
df_model_pref_srl_mag_coef['model_srl1_fxd (preferred)', 'sigma'] <- sigma(fit_model_srl1m_fxd)
df_model_pref_srl_mag_coef['model_srl3p_fxd (alternative)','sigma'] <- sigma(fit_model_srl3pm_fxd)
#rename columns
colnames(df_model_pref_srl_mag_coef)[colnames(df_model_pref_srl_mag_coef)=='(Intercept)'] <- 'intrcp'
colnames(df_model_pref_srl_mag_coef)[colnames(df_model_pref_srl_mag_coef)=='x1'] <- 'X1'
colnames(df_model_pref_srl_mag_coef)[colnames(df_model_pref_srl_mag_coef)=='x3prime'] <- 'X3prime'
#reorder model coefficients
df_model_pref_srl_mag_coef <- df_model_pref_srl_mag_coef[,c('intrcp','X1','X3prime','FD','sigma')]
## Comparison of different models
fprintf("SRL Modeling:
\tModel1 fxd. (preferred):\n\t\tLogLik: %.2f\n\t\tAIC: %.2f\n\n
\tModel3' fxd. (alternative):\n\t\tLogLik: %.2f\n\t\tAIC: %.2f\n\n",
logLik(fit_model_srl1m_fxd), AIC(fit_model_srl1m_fxd), logLik(fit_model_srl3pm_fxd), AIC(fit_model_srl3pm_fxd) )
## SRL Modeling:
## Model1 fxd. (preferred):
## LogLik: 14.49
## AIC: -22.99
##
##
## Model3' fxd. (alternative):
## LogLik: 12.05
## AIC: -18.10
#empirical sigma below and above width threshold of model 1
sig_m1_fxd_wu <- sqrt(mean(df_FDHI_srl[!df_FDHI_srl$flag_wlim,'srl1m_fxd_res']^2))
sig_m1_fxd_wl <- sqrt(mean(df_FDHI_srl[ df_FDHI_srl$flag_wlim,'srl1m_fxd_res']^2))
sig_m3p_fxd_wu <- sqrt(mean(df_FDHI_srl[!df_FDHI_srl$flag_wlim,'srl3pm_fxd_res']^2))
sig_m3p_fxd_wl <- sqrt(mean(df_FDHI_srl[ df_FDHI_srl$flag_wlim,'srl3pm_fxd_res']^2))
fprintf("Standard Deviations:\n\tModel 1 fxd. (preferred):\n\t\tEmpirical ratio for std of width limted to unbounded events is: %.2f\n
Standard Deviations:\n\tModel 3' fxd. (alternative):\n\t\tEmpirical ratio for std of width limted to unbounded events is: %.2f\n\n",
sig_m1_fxd_wu/sig_m1_fxd_wl, sig_m3p_fxd_wu/sig_m3p_fxd_wl)
## Standard Deviations:
## Model 1 fxd. (preferred):
## Empirical ratio for std of width limted to unbounded events is: 1.92
##
## Standard Deviations:
## Model 3' fxd. (alternative):
## Empirical ratio for std of width limted to unbounded events is: 1.77
| Intercept | \(X_1\) scaling | \(X_3'\) scaling | \(F_D\) | \(\sigma\) | |
|---|---|---|---|---|---|
| model_srl1_fxd (preferred) | -4.172824 | 1 | NA | -0.1286700 | 0.2502627 |
| model_srl3p_fxd (alternative) | -4.111267 | NA | 1 | -0.1855579 | 0.2553302 |
Regression summary of model 1_fxd:
##
## Call:
## lm(formula = form_model_srl1m_fxd, data = df_FDHI_srl, weights = reg_wt)
##
## Weighted Residuals:
## Min 1Q Median 3Q Max
## -0.63369 -0.17869 0.02732 0.16560 0.56149
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -4.17282 0.02429 -171.761 < 2e-16 ***
## FD -0.12867 0.04253 -3.025 0.00304 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.2503 on 120 degrees of freedom
## Multiple R-squared: 0.07086, Adjusted R-squared: 0.06312
## F-statistic: 9.152 on 1 and 120 DF, p-value: 0.00304
Regression summary of model 3':
##
## Call:
## lm(formula = form_model_srl3pm_fxd, data = df_FDHI_srl, weights = reg_wt)
##
## Weighted Residuals:
## Min 1Q Median 3Q Max
## -0.71561 -0.17854 0.01277 0.15730 0.55563
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -4.11127 0.02479 -165.869 < 2e-16 ***
## FD -0.18556 0.04339 -4.276 3.84e-05 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.2553 on 120 degrees of freedom
## Multiple R-squared: 0.1322, Adjusted R-squared: 0.125
## F-statistic: 18.29 on 1 and 120 DF, p-value: 3.838e-05
Figure 6 presents the predictions of all candidate \(SRL\) models for Strike-slip events with a
\(15 km\) and \(20 km\) width of seismogenic zone. Figure 7
presents the predictions for model 1 and
model 1 fxd, Figure 8 presents the predictions for
model 2 and model 2', and Figure 9 presents
the predictions for model 3, model 3', and
model 3' fxd.
Figure 6. Prediction of candidate SRL Models for Strike-slip events with a 15km and 20km width of seismogenic zone
Figure 7. Prediction of SRL model 0 for Strike-slip events
Figure 8. Prediction of SRL model 1 for Strike-slip events with a 15km and 20km width of seismogenic zone
Figure 9. Prediction of SRL model 2 for Strike-slip events with a 15km and 20km width of seismogenic zone
Figure 10. Prediction of SRL model 3 for Strike-slip events with a 15km and 20km width of seismogenic zone
Figures 11 and 12 present the predictions of model 1 fxd
(preferred model) and model 3' fxd (alternative model) for
Strike-slip events with a \(15 km\) and
\(20 km\) width of seismogenic zone,
respectively.
Figure 11. Prediction of preferred SRL Model for Strike-slip events with a 15km and 20km width of seismogenic zone
Figure 12. Prediction of alternative SRL Model for Strike-slip events with a 15km and 20km width of seismogenic zone
Figures 13 and 14 present the predictions of model 1 fxd
(preferred model) and model 3' fxd (alternative model) for
Strike-slip and Dip-slip events with a \(15
km\) width of seismogenic zone.
Figure 13. Prediction of preferred SRL Model for Strike-slip and Dip-slip events and 15km width of seismogenic zone
Figure 14. Prediction of alternative SRL Model for Strike-slip and Dip-slip events and 15km width of seismogenic zone
Figure 15 compares the predictions of model 0,
model 1 fxd, and model' 3 fxd for Strike-slip
events with a \(15 km\) and \(20 km\) width of seismogenic zone. All
models have similar scaling for unbounded ruptures, but
model 1 fxd and model' 3 fxd also capture the
change in scaling for width-limited ruptures leading to an overall
better predictive performance.
Figure 15. Prediction of Candidate SRL Models for Strike-slip events with a 15km and 20km width of seismogenic zone. Comparison with unbounded and width-limited events
Figure 16 shows the model 0 regression residuals versus
magnitude.
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
Figure 16. Linear model (Model 0, WC type) Residuals versus Magnitude
Figure 17 shows the regression residuals versus magnitude of
model 1 fxd (preferred model).
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
Figure 17. Prefered Model (Model 1, D~sqrt(A), fixed slope) Residuals versus Magnitude
Figure 18 shows the regression residuals versus magnitude of
model 3' fxd (alternative model).
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
Figure 18. Alternative Model (Model 3’, D~sqrt(A) - L, fixed slope) Residuals versus Magnitude
Figure 19 to 21 show regression residuals versus magnitude for
model 1, model 2, and
model 3.
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
Figure 19. Candidate Model (Model 1, D~sqrt(A)) Residuals versus Magnitude
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
Figure 20. Candidate Model (Model 2, D~W) Residuals versus Magnitude
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
Figure 21. Candidate Model (Model 3, D~L) Residuals versus Magnitude
Figure 22 and 23 show regression residuals versus magnitude for
model 2' and model 3'.
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
Figure 22. Candidate Model (Model 2’, D~sqrt(A)-L) Residuals versus Magnitude
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
Figure 23. Candidate Model (Model 3’, D~sqrt(A)-L) Residuals versus Magnitude
A rupture-width transition parameter \(\delta\) is defined as \(\delta = \log_{10}(W_{unb}) -
\log_{10}(W_{lim})\); \(W_{unb}\) is the theoretical rupture width
for an infinite seismogenic zone estimated from the first leg of
model w1 (\(\log_{10}(W_{unb}) =
\beta_1 \mathbf{M} + \beta_2\)). A negative value of \(\delta\) corresponds to a unbounded
rupture, whereas a positive corresponds to a width-limited rupture.
Figure 24 and 25 show the regression residuals versus \(\delta\) for the preferred
(model 1 fxd) and alternative (model 3' fxd)
models. The solid black lines correspond to the rupture-width transition
dependent standard deviation.
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
Figure 24 Prefered Model (Model 1, D~sqrt(A), fixed slope) Residuals versus Width transition
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
Figure 25 Alternative Model (Model 3’, D~sqrt(A) - L, fixed slope) Residuals versus Width transition
## Warning: Removed 1 rows containing missing values (`geom_pointrange()`).
## Removed 1 rows containing missing values (`geom_pointrange()`).
## Removed 1 rows containing missing values (`geom_pointrange()`).
## Removed 1 rows containing missing values (`geom_pointrange()`).
## Removed 1 rows containing missing values (`geom_pointrange()`).
Figure 23 compares the preferred model with existing \(SLR\) relationships. WC94 All and WC94 SS correspond to the Wells and Coppersmith (1994) \(SRL ~ \mathbf{M}\) scaling relationships for All types of slip and Strike-slip events. W08 All, W08 SS, and W08 R correspond to the Wesnousky (2008) \(SRL ~ \mathbf{M}\) scaling relationships for All, Dip-slip, Strike-slip, and Reverse Shallow-Crustal events. L10 DS, L10 SS, and L10 SRC correspond to the Leonard (2010) \(SRL ~ \mathbf{M}\) scaling relationships for Dip-slip, Strike-slip, and Stable-Continental events. L14 Interpalte SS and L14 Interpalte DS correspond to the Leonard (2014) \(SRL ~ \mathbf{M}\) scaling relationships for Strike-slip and Dip-slip Interplate events. L14 Interpalte SS and L14 Interpalte DS correspond to the Leonard (2014) \(SRL ~ \mathbf{M}\) scaling relationships for Strike-slip and Dip-slip Interplate events, while L14 SCR DS corresponds to Dip-slip Stable-Continental events. WY15 OLS and WY15 Deming correspond to the Wells and Youngs (2015) \(SRL ~ \mathbf{M}\) scaling relationships using ordinary-least-squares and errors-in-variables regression models. The proposed model is in agreement with the existing models over different magnitude ranges. It is in good agreement with the Leonard (2010) scaling relationship for small to moderate size events, and in good agreement with the Wells and Coppersmith (1994) and Wells and Youngs (2015) scaling relationships at large size events. A potential reason for the agreement at different magnitude ranges is that the dataset used by Leonard (2010) was primarily composed of events ranging from \(\mathbf{M}~5\) to \(7\), while datasets used by Wells and Coppersmith (1994) and Wells and Youngs (2015) was primarily composed of events ranging from \(\mathbf{M}~6\) to \(7.8\).
## Warning: Removed 13 rows containing missing values (`geom_line()`).
## Warning: Removed 26 rows containing missing values (`geom_line()`).
## Warning: Removed 6 rows containing missing values (`geom_line()`).
## Warning: Removed 16 rows containing missing values (`geom_line()`).
## Warning: Removed 3 rows containing missing values (`geom_line()`).
## Warning: Removed 4 rows containing missing values (`geom_line()`).
## Warning: Removed 13 rows containing missing values (`geom_line()`).
## Warning: Removed 26 rows containing missing values (`geom_line()`).
## Warning: Removed 6 rows containing missing values (`geom_line()`).
## Warning: Removed 16 rows containing missing values (`geom_line()`).
## Warning: Removed 3 rows containing missing values (`geom_line()`).
Figure 26. Comparison Selected Models (Model 1, D~sqrt(A), fixed slope) & (Model 3’, D~sqrt(A) - L, fixed slope) with existing Active Shallow Crustal Strike-Slip SLR relationships
Figure 27. Comparison Selected Models (Model 1, D~sqrt(A), fixed slope) & (Model 3’, D~sqrt(A) - L, fixed slope) with existing Active Shallow Crustal Dip-Slip SLR relationships
## Warning: Removed 4 rows containing missing values (`geom_line()`).
Figure 28. Comparison Selected Models (Model 1, D~sqrt(A), fixed slope) & (Model 3’, D~sqrt(A) - L, fixed slope) with existing Stable Continental Region SLR relationships
## Warning: Removed 13 rows containing missing values (`geom_line()`).
## Warning: Removed 26 rows containing missing values (`geom_line()`).
## Warning: Removed 60 rows containing missing values (`geom_line()`).
## Warning: Removed 4 rows containing missing values (`geom_line()`).
## Warning: Removed 6 rows containing missing values (`geom_line()`).
## Warning: Removed 16 rows containing missing values (`geom_line()`).
## Warning: Removed 3 rows containing missing values (`geom_line()`).
Figure 29. Comparison Selected Models (Model 1, D~sqrt(A), fixed slope) & (Model 3’, D~sqrt(A) - L, fixed slope) with existing SLR relationships
Baize, S., O. Scotti, J.-F. Ritz, M. Ferry, C. Larroque, E. Mathot, L. Audin, F. Nurminen, P. Boncio, and B. Delouis (2020). Surface ruptures during moderate earthquakes: Is that so rare? Geological Society of America 2020 Connects Online
Bodin, P., and J. N. Brune (1996). On the scaling of slip with rupture length for shallow strike-slip earthquakes: Quasistatic models and dynamic rupture propagation, Bull. Seismol. Soc. Am. 86, 1292.
Hanks, T. C., & Bakun, W. H. (2002). A bilinear source-scaling model for M-log A observations of continental earthquakes. Bulletin of the Seismological Society of America, 92(5), 1841-1846.
Leonard, M. (2010). Earthquake fault scaling: Self-consistent relating of rupture length, width, average displacement, and moment release. Bulletin of the Seismological Society of America, 100(5A), 1971-1988.
Leonard, M. (2014). Self‐consistent earthquake fault‐scaling relations: Update and extension to stable continental strike‐slip faults. Bulletin of the Seismological Society of America, 104(6), 2953-2965.
Romanowicz, B. (1992). Strike‐slip earthquakes on quasi‐vertical transcurrent faults: Inferences for general scaling relations. Geophysical Research Letters, 19(5), 481-484.
Romanowicz, B. (1994). Comment on “A reappraisal of large earthquake scaling” by C. Scholz. Bulletin of the Seismological Society of America, 84(5), 1675-1676.
Romanowicz, B., & Rundle, J. B. (1993). On scaling relations for large earthquakes. Bulletin of the Seismological Society of America, 83(4), 1294-1297.
Pegler, G., & Das, S. (1996). Analysis of the relationship between seismic moment and fault length for large crustal strike‐slip earthquakes between 1977–92. Geophysical research letters, 23(9), 905-908.
Scholz, C. H. (1982). Scaling laws for large earthquakes: consequences for physical models. Bulletin of the Seismological Society of America, 72(1), 1-14.
Scholz, C. H. (1994). A reappraisal of large earthquake scaling. Bulletin of the Seismological Society of America, 84(1), 215-218.
Scholz, C. H. (1997). Size distributions for large and small earthquakes. Bulletin of the Seismological Society of America, 87(4), 1074-1077.
Scholz, C. H. (1998). A further note on earthquake size distributions. Bulletin of the Seismological Society of America, 88(5), 1325-1326.
Wang, Y., & Goulet, C. (2021). Validation of fault displacements from dynamic rupture simulations against the observations from the 1992 landers earthquake. Bulletin of the Seismological Society of America, 111(5), 2574-2594.
Wells, D. L., & Coppersmith, K. J. (1994). New empirical relationships among magnitude, rupture length, rupture width, rupture area, and surface displacement. Bulletin of the seismological Society of America, 84(4), 974-1002.
Wells, D. L., & Youngs R. B. (2015). Improved Regression Relations for Earthquake Source Parameters. Seismological Society of America 2015 Annual Meeting
Wesnousky, S. G. (2008). Displacement and geometrical characteristics of earthquake surface ruptures: Issues and implications for seismic-hazard analysis and the process of earthquake rupture. Bulletin of the Seismological Society of America, 98(4), 1609-1632.